SynForecast Roadmap

Augmentation

Optional moment matching — augmented series are pinned to the source mean, std, and lag-1 ACF by construction; add match_moments=False and test whether unpinned should be the default. Highest priority here.
Strategy selection — ~~unified API~~ — done: fit-and-simulate, mixup, MBB and DBA ship as frozen specs with apply(). Still open: the jitter/scale/warp baselines.
Wider fitter coverage — 9 fitters against 32 generators, so only a third of the library is reachable; add ETS, StateSpace, INAR, multi-period seasonal.
Panel-aware augmentation — augment jointly to preserve cross-series correlation instead of series-by-series.
~~Exogenous passthrough~~ — done: extra columns are carried through augmentation.
X-conditioned fit-and-simulate — simulate the target driven by given exogenous inputs.
Fit diagnostics — return ranked candidates with fit scores so a poor best-fit is visible rather than silent.
Rust batch path — route fitted-generator simulation through it so augmentation scales like generation.

Feature-space coverage — shipped in synforecast.evaluation: a frozen twelve-feature native_v1 schema and real-anchored PCA grid coverage. The benchmarks score coverage instead by full-feature nearest-neighbour distance at a radius calibrated on held-out real series; promote that score to the API so users can reproduce them. Then add tsfeatures/catch22 parity and per-feature attribution of uncovered regions. Benchmark scripts follow separately.
Seasonal coverage — seasonal_pool is opt-in. Next: measure the forecasting utility of pool + seasonal_pool pretraining on a seasonal panel, and run a pre-declared test of a period-aware default (on only for periods 4 and 12).
Nearest-neighbour distance to real data — promote the memorization check from notebook to API; it shares the machinery of calibrated coverage above.
Fidelity scores — discriminative and predictive scores for comparing generation methods.
Wider benchmarks — extend when_synthetic_helps beyond one model and dataset.

Rust SARIMA exogenous support, then drop the Python SARIMA path.
Rust batch dispatch for Copula, VAR, DailyActiveUsers, IoTSensor.
Streaming generation — generate_values and members= give DataFrame-free, per-draw generation in memory; next is chunked Parquet/Arrow output for corpora larger than memory.
TCM cost — TCM takes over 80% of pretraining_pool generation time at every length from 128 to 8,192 steps, with 3 of about 60 entries; speed up its causal-model simulation.
Fresh-seed draws — rebuilding a pool member with a new seed costs more than drawing its series; make per-draw reseeding cheaper than a full rebuild.
Long-series KernelSynth — sampling is O(n^2) per kernel term but only about 11% of pool time at 8,192 steps; sample fast-decaying terms by circulant embedding if corpora get longer.

Hierarchical panels — panels that aggregate coherently, for testing reconciliation.
Generator selection guide — a decision path to a generator or pool, instead of reading all 32 pages.

Deep generative models (GAN/VAE/diffusion) — heavy deps, GPU time, and per-dataset training conflict with cheap deterministic CPU generation;
Irregular sampling — non-uniform time grids and channels observed at different times.